Anti-tailgating detection method based on video recognition

By compressing the time axis to the channel dimension in the video recognition network and using the time channel learning module, combined with pre-training and fine-tuning methods, the problem of blind spots and motion blur in video anti-trailing detection is solved, and efficient and accurate recognition is achieved. And reduces the cost of computing power.

CN118711115BActive Publication Date: 2024-11-29SHENZHEN MAXVISION TECH
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411207985.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-30
Publication Date
2024-11-29
Estimated Expiration
2044-08-30

AI Technical Summary

Technical Problem

The existing video anti-trailing algorithm has error detection problems when dealing with blind spots and motion blur of view angles. The video recognition method based on 3D convolution or 2D convolution requires a large amount of computing resources and is difficult to apply on mobile devices. And due to the relatively high trailing cases, it is not possible to use. It is difficult to obtain enough training data, resulting in high computing power costs.

Method used

Using a lightweight video recognition network, the time axis of the video sequence is compressed into the channel dimension, and the time position dynamics and time object relationships of the video image sequence are learned through the time channel learning module (CTL), avoiding additional time axis modeling operations. , combined with pre-training and fine-tuning methods, solve the problem of scarcity of labeled data.

Benefits of technology

The model calculation volume is reduced, the recognition accuracy is improved, the fast video understanding is achieved on mobile devices, and the performance of downstream tasks is improved through pre-training models to adapt to specific tasks on small data sets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118711115B_ABST
    Figure CN118711115B_ABST
Patent Text Reader

Abstract

The invention discloses an anti-tailing detection method based on video recognition, comprising the following steps: step A: inputting a video image frame sequence, and adopting an equally spaced sampling strategy to intercept the video image; step B: during data preprocessing, converting the video image frame sequence into an input format required by a video recognition network; step C: constructing a video recognition network model to realize accurate recognition and real-time prediction of the model; step D: defining a classification loss function to supervise model training, wherein the model gradually reaches a convergence state through iteration of weight parameters during the training process, and during the training period, selecting the optimal network weight parameters in a verification set for storage, and generating a video recognition model; step E: in the actual reasoning process, directly loading the weight parameters of the video recognition network model to complete the reasoning process of the model, and inputting the preprocessed data into the video recognition model for prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of anti-tailgating detection, and in particular to an anti-tailgating detection method based on video recognition. Background Art

[0002] Existing video anti-tailing algorithms are usually based on target detection methods, which use deep learning target detection methods to detect pedestrians or heads in images captured by cameras, and judge whether there is tailing in the monitoring area by the number of heads or pedestrians. Ordinary cameras inevitably have blind spots and cannot cope with some special tailing situations, such as children whose bodies are blocked by adults, babies in arms, etc. Moreover, when pedestrians move faster, the image frames captured often have motion blur, which is another major challenge for the detection algorithm and often results in false detections. Methods based on video recognition extract spatiotemporal features from image sequences, combining temporal and spatial features to detect pedestrians. The current video recognition architecture is mostly based on 3D convolution blocks or 2D convolutions with additional time modeling operations. The time axis needs to be regarded as an independent dimension in the video sequence, which requires a lot of computing and memory resources and limits its application on mobile devices. In addition, when the video recognition method is applied to anti-tailing scenarios, a large number of normal and tailing videos need to be obtained for training. However, in actual applications, tailing cases are occasional phenomena, and it is difficult to obtain a large number of videos as training sets. In addition, the annotation of video data requires a lot of time and effort, and the computing cost is relatively high. Summary of the invention

[0003] An object of the present invention is to provide an anti-tailgating detection method based on video recognition, which divides input video data into different categories or labels through video recognition, so that the video content can be automatically understood and recognized, and the intelligent classification and retrieval of the video content can be realized. In addition, the method based on video recognition can extract spatiotemporal features of image sequences, make predictions based on comprehensive temporal and spatial feature information, and avoid the problems of visual blind spots and motion blur. Moreover, through a lightweight video recognition network, the time axis of the video sequence is compressed into the channel dimension, and a temporal channel learning module (CTL) is used to help the model learn the temporal position dynamics and temporal object relationships of the video image sequence, eliminating the need for additional modeling operations on the time axis, greatly reducing the amount of model operations and improving the recognition accuracy, so that when the network runs on a mobile device, it can achieve a faster video understanding speed while maintaining a fairly high accuracy.

[0004] Other advantages and features of the present invention will become more apparent from the following detailed description and may be realized by means of the instrumentalities and combinations particularly pointed out in the appended claims.

[0005] According to one aspect of the present invention, a video recognition-based anti-tailgating detection method of the present invention that can achieve the aforementioned objectives and other objectives and advantages comprises the following steps:

[0006] Step A: Input a video image frame sequence and capture the video image using an equally spaced sampling strategy;

[0007] Step B: During data preprocessing, the video image frame sequence is converted into the input format required by the video recognition network;

[0008] Step C: Build a video recognition network model to achieve accurate recognition and real-time prediction of the model;

[0009] Step D: Define the classification loss function to supervise the model training. During the training process, the model gradually reaches a convergence state through the iteration of weight parameters. During the training, the optimal network weight parameters in the validation set are selected and saved to generate a video recognition model.

[0010] Step E: In the actual reasoning process, directly load the weight parameters of the video recognition network model to complete the reasoning process of the model. In the reasoning mode, sample the input anti-tailing surveillance video to obtain a video image frame sequence, perform image preprocessing, and input the preprocessed data into the video recognition model for prediction. If the recognition result is normal, no alarm is triggered and the video stream continues to be input for monitoring. If the recognition result is tailing, an alarm signal is triggered.

[0011] According to an embodiment of the present invention, step A includes sub-step A1: capturing a frame of image at a fixed frame interval, where the frame interval ranges from 4 to 8, and the t-frame image sequence input to the video recognition network is recorded as .

[0012] According to an embodiment of the present invention, the step B includes sub-step B1: converting the input format includes resizing, standardizing, and dimensional conversion.

[0013] According to one embodiment of the present invention, step C includes sub-step C1: the video recognition network is lightweight and adopts time compression. The video recognition network includes a dimensionality conversion module, a convolution module, a maximum pooling module, four CTL modules, a downsampling module, an average pooling module and a fully connected module.

[0014] According to one embodiment of the present invention, step C includes sub-step C2: the CTL module includes two branches: one branch is responsible for learning the temporal importance, and the other branch with temporal position recovery capability can enhance the object modeling capability across time.

[0015] According to an embodiment of the present invention, step C includes sub-step C3: the CTL module is a residual structure, including a 1x1 convolution, a CTL module and a 1x1 convolution,

[0016] According to an embodiment of the present invention, step C includes sub-step C4: the CTL module is a dual-branch structure, and the output of the CTL module is the sum of the output feature vectors of the two branches.

[0017] According to one embodiment of the present invention, step C includes sub-step C5: the first branch of the CTL module is composed of TFC, i.e., temporal focused convolution, TFC introduces an attention mechanism to calculate the weight of each feature channel to obtain a feature vector with temporal importance, and the second branch of the CTL module is composed of IOI cross-time interaction, and IOI restores the temporal position dynamics and temporal object relationships.

[0018] According to one embodiment of the present invention, step C includes sub-step C6: the IOI module is composed of two branches, one branch uses TFC to reduce the number of channels C to the number of frames T, while capturing the temporal importance, injecting temporal position coding information, restoring temporal dynamics, using 7 × 7 convolution to model the object relationship between T frames, using 3 × 3 convolution to map the number of channels from T to C, using an activation function to obtain a weight vector in the two-dimensional space of time and position, and the other branch uses 3 × 3 convolution to output a feature map with a channel of C, performs corresponding element multiplication operations with the weight vector, and obtains a feature vector with temporal position dynamics and temporal object relationship.

[0019] According to one embodiment of the present invention, step D includes sub-step D1: in training mode, each training video is parsed into an image sequence and labeling is performed. The labels are divided into two categories: normal and trailing. The labeling rule is: when there is no person or only one person in the video, it is labeled as normal; when the number of people in the video is greater than one person, it is labeled as trailing.

[0020] The beneficial effects of the present invention are as follows: through a lightweight video recognition network, the time axis of the video sequence is compressed into the channel dimension, and the time channel learning module (CTL) is used to help the model learn the time position dynamics and time object relationship of the video image sequence, thereby eliminating the need for additional modeling operations on the time axis, greatly reducing the amount of model operations and improving the recognition accuracy, so that when the network is running on a mobile device, it can achieve a faster video understanding speed while maintaining a fairly high accuracy rate. In addition, the method of deep learning pre-training plus fine-tuning can solve the problem of scarce labeled data, pre-train the model on a large data set in a supervised or unsupervised manner, and adapt the pre-trained model to specific downstream tasks on a smaller data set through fine-tuning, which can avoid training new models from scratch for different tasks or data sets. Pre-training on a larger data set helps to learn general representations, thereby improving the performance of downstream tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The present invention is a flowchart of an anti-tailgating detection method based on video recognition according to an embodiment of the present invention.

[0022] Figure 2 This is a diagram of a time-compressed video recognition network structure of the anti-tailgating detection method based on video recognition in the above embodiment of the present invention.

[0023] Figure 3 It is a structural diagram of a CTL module of the anti-tailgating detection method based on video recognition according to the above embodiment of the present invention.

[0024] Figure 4 It is a structural diagram of a CTL module of the anti-tailgating detection method based on video recognition according to the above embodiment of the present invention.

[0025] Figure 5 It is a TFC network structure diagram of the anti-tailgating detection method based on video recognition according to the above embodiment of the present invention.

[0026] Figure 6 1 is an IOI network structure diagram of the anti-tailgating detection method based on video recognition according to the above embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art can think of other obvious variations. The basic principles of the present invention defined in the following description can be applied to other embodiments, variations, improvements, equivalents, and other technical solutions that do not deviate from the spirit and scope of the present invention.

[0028] Those skilled in the art should understand that, in the disclosure of the present invention, the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicating the orientation or position relationship are based on the orientation or position relationship shown in the drawings, which are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the above terms should not be understood as limiting the present invention.

[0029] It is understood that the term "one" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "one" cannot be understood as a limitation on the number. In addition, in the description of the present invention, unless otherwise specified, the meaning of "multiple" is two or more than two.

[0030] In the description of the present invention, it should be understood that "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through a medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0031] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0032] like Figures 1 to 6As shown, an anti-tailing detection method based on video recognition of an embodiment of the present invention is disclosed. The anti-tailing detection method based on video recognition can classify the video of the monitoring area collected by the camera into two categories: normal and tailing. The video category is learned through the model, and then the input video of unknown category is predicted to determine whether there is tailing in the current monitoring area. The anti-tailing detection method based on video recognition includes the following steps:

[0033] Step 1: Input the video image frame sequence. In view of the temporal redundancy in the continuous video frames, the video images are intercepted by an equal interval sampling strategy, that is, one frame of image is intercepted at a fixed interval of frames. Preferably, the interval frame number ranges from 4 to 8. The t-frame image sequence input to the video recognition network is recorded as ;

[0034] Step 2: During data preprocessing, the video image frame sequence is converted into the input format required by the video recognition network, including but not limited to size adjustment, standardization, dimension conversion, etc.;

[0035] Step 3: Build a video recognition network model to achieve accurate recognition and real-time prediction of the model, so that the entire network model is streamlined and efficient. In detail, when building the network model, it is preferred to use a lightweight video recognition network, that is, a time-compressed video recognition network, such as Figure 2 As shown in the figure, it is a temporal compressed video network structure diagram. The specific approach is as follows: the time axis of the video image frame sequence is compressed into the channel dimension. In order to enhance the temporal modeling ability of the network, a channel time learning (CTL) module is designed to capture the temporal dynamics of the video image frame sequence. The module contains two complementary branches, one of which is responsible for learning the temporal importance, and the other branch with temporal position recovery ability can enhance the cross-temporal object modeling ability.

[0036] The network consists of multiple stages, four of which are stacked by CTL modules (CTL Block), which can mine and recover hidden time representations. It is worth noting that the network includes a dimension conversion module, a convolution module, a maximum pooling module, four CTL modules, a downsampling module, an average pooling module and a fully connected module. Among them, the dimension conversion module, the convolution module, the maximum pooling module, the average pooling module and the fully connected module are all existing technologies and will not be described in detail here.

[0037] In short, the dimension conversion module converts the dimension of the input image vector. For example, if the input dimension is (1, 3, t, h, w), the time dimension is compressed into the channel after the dimension conversion module is operated, and the output dimension is (1, 3*t, h, w), where 1 is the batch number of the input image, that is, the image vector of one batch is input, t is the number of frames of the input image sequence, h and w represent the width and height of the image respectively, 3 is the number of channels of the image,

[0038] Among them, the (1,3*t,h,w) image vector is obtained through the 5*5 convolution module, which is still the basic feature compressed in the time dimension. The dimension is (1,c1,h / 2,w / 2), where c1 is the number of output channels of the 5*5 convolution module. The time dimension and channel order of the basic feature are disrupted. Next, the time position dynamics and time object relationship are restored through four stages of CTL modules. Downsampling is used before each CTL module to reduce the size of the intermediate feature map and reduce the amount of network calculation. The basic feature (1,c1,h / 2,w / 2) is input into the maximum pooling module, and the output dimension is (1,c1,h / 4, w / 4), the pooled feature (1,c1,h / 4,w / 4) is input into the first CTL module, and the output dimension is (1,c2,h / 4,w / 4). The feature (1,c2,h / 4,w / 4) is input into the first convolution downsampling module with a convolution kernel of 2*2 and a stride of 2, and the output dimension is (1,c2,h / 8,w / 8). The feature (1,c2,h / 8,w / 8) is input into the second CTL module, and the output dimension is (1,c3,h / 8,w / 8). The feature (1,c3,h / 8,w / 8) is input into the first convolution downsampling module with a convolution kernel of 2*2 and a stride of 2. The second convolution downsampling module outputs features of dimension (1, c3, h / 16, w / 16). The feature (1, c3, h / 16, w / 16) is input into the third CTL module, and the output dimension is (1, c4, h / 16, w / 16). The feature (1, c4, h / 16, w / 16) is input into the third convolution downsampling module with a convolution kernel of 2*2 and a stride of 2, and the output dimension is (1, c4, h / 32, w / 32). The feature (1, c4, h / 32, w / 32) is input into the fourth CTL module, and the output dimension is (1, c5, h / 32, w / 3 2), the feature (1, c5, h / 32, w / 32) is input into the average pooling module, and the output dimension is (1, c5, h / 64, w / 64). The fully connected module is used to process the extracted convolutional features and generate the final output. The learned distributed feature representation is mapped to the sample label space. The average pooling feature (1, c5, h / 64, w / 64) is input into the fully connected module to obtain the predicted category vector. The activation function converts the category vector into a classification probability vector with a numerical range of 0~1. The loss function is then connected to generate a complete training network. The loss function is used to supervise the weight training of the entire network.

[0039] Furthermore, the CTL module is used as the core module of the temporally compressed video recognition network, such as Figure 3 As shown in the figure, it is the structure diagram of the CTL module. The CTL module is a residual structure, including 1x1 convolution, CTL module and 1x1 convolution. Figure 4 As shown in the figure, it is a CTL module structure diagram. The CTL module has a dual-branch structure. The output of the CTL module is the sum of the output feature vectors of the two branches.

[0040] The first branch of the CTL module is composed of TFC (Temporal Focus Convolution), which is used to calculate the time importance. As an improved version of ordinary convolution, the time importance of each channel is different for the time-compressed feature vector. Considering that ordinary convolution regards different channels as of equal importance when aggregating channels, it cannot express the time importance of each channel well. Therefore, the TFC of the first branch introduces an attention mechanism to calculate the weight of each feature channel. Before the ordinary convolution operation, different weights are given to the features of each channel, so as to better express the time importance of each channel feature. Figure 5 As shown in the figure, it is the network structure diagram of TFC, where the right branch is the attention branch. The Sigmoid activation of the attention branch outputs the weights of each channel feature, performs corresponding element multiplication operation with the input feature vector on the left, and then performs 1*1 convolution calculation to obtain a feature vector with temporal importance.

[0041] The second branch of the CTL module is composed of IOI (Inter-temporal Object Interaction). Figure 6 As shown in Figure 1, it is the network structure diagram of IOI. The setting of IOI branch can solve the following two problems: on the one hand, when compressing the time information of the anti-tailing video image frame sequence into the channel, after being processed by the multi-layer neural network, the time order information of the channel will be confused and some discriminative details will be lost. On the other hand, for time-compressed video clips, a moving object may appear in different channels over time. Therefore, it is important to capture the relationship between targets attached to different time channels. The model needs to learn spatial relationships, and the IOI branch can restore the temporal position dynamics and temporal object relationships, which just solves the above two problems. Furthermore, the IOI module consists of two branches. The right branch first uses TFC to reduce the number of channels (C) to the number of frames (T), while capturing the temporal importance, injecting temporal position encoding information, and restoring temporal dynamics. The object relationship between T frames is modeled using 7 × 7 convolution, and the number of channels is mapped from T to C using 3 × 3 convolution. The weight vector in the two-dimensional space of time and position is obtained using the activation function. The left branch uses 3 × 3. The convolution output channel is the feature map of C, which is multiplied by the corresponding element with the weight vector on the right to obtain a feature vector with time position dynamics and time object relationship;

[0042] Step 4: Define the classification loss function to supervise the model training. The model gradually reaches the convergence state through the iteration of weight parameters during the training process. During the training, the optimal network weight parameters in the verification set are selected and saved to generate a video recognition model. In the training mode, each training video is parsed into an image sequence and labeled. The labels are divided into two categories: normal and trailing. The labeling rule is: when there is no one or only one person in the video, it is marked as normal; when there are more than one person in the video, it is marked as trailing;

[0043] Step 5: In the actual reasoning process, directly load the weight parameters of the video recognition network model to complete the reasoning process of the model. In the reasoning mode, sample the input anti-tailing surveillance video to obtain a video image frame sequence, perform image preprocessing, and input the preprocessed data into the video recognition model for prediction. If the recognition result is normal, no alarm is triggered and the video stream continues to be input for monitoring. If the recognition result is tailing, an alarm signal is triggered.

[0044] In actual application scenarios, the specific implementation steps of the anti-tailing detection method based on video recognition in this solution are as follows:

[0045] First, we collect public video resources for network self-supervision training. The backbone network is a video recognition network based on time compression. Considering the difficulty of obtaining balanced anti-tailing videos as training data, we adopt the method of pre-training plus fine-tuning to avoid overfitting caused by too little training data, so that even with less data, we can achieve good generalization performance. There are abundant unlabeled video data resources on the Internet: there are a large number of public video data sets and short video free download websites, which can collect millions or even tens of millions of training data. When the training data is huge, the self-supervised training process takes a long time. For example, when using 16 3090 Nvidia cards, the training cycle is usually 25-30 days. Among them, self-supervised learning is a machine learning method, and its model can learn useful representations from unlabeled data. The key to the method is to design one or more self-supervised tasks, which do not require manually labeled data, but use the intrinsic structure or attributes of the data to generate training signals. In this way, even if these tasks themselves do not directly correspond to the final application scenario, the model can learn feature representations that are useful for downstream tasks. The training principle of the self-supervised model also includes learning the intrinsic laws and feature representations of the data through self-supervised tasks, which can be predicting the damaged parts of the image, segmenting different parts from the image, or representing the image as a visual token through techniques such as variational autoencoders. In this way, the model can learn feature representations that are useful for visual or language tasks, thereby improving performance on downstream tasks.

[0046] In the second step, two types of videos, normal and trailing, are collected in the application scenario. A 16-frame image sequence is captured for each video using a 4-frame interval sampling method, and the category labels are annotated to generate training and test data respectively. The self-supervised model trained in the first step is used as a pre-training model, and the softmax cross entropy loss function is used for model fine-tuning training. The trained weight parameter file is saved.

[0047] The third step is to convert the model file stored in the second step into the format file required by the model application platform, and deploy the application. The video stream collected by the anti-tailgating device is sent to the model calculation platform. The platform parses the video into an image sequence according to the 4-frame interval sampling method. Every 16 frames of images are input into the model as a data packet, and the prediction results are output and sent to the anti-tailgating warning platform. The warning platform issues an alarm or a pass command based on the prediction results.

[0048] The anti-tailing detection method based on video recognition in the above-mentioned embodiment of the present invention can effectively overcome the two major problems of visual blind spots and motion blur. The video classification method does not need to judge the tailing result based on the number of detected targets, but only uses the spatiotemporal information of a multi-frame image sequence to model and analyze the video content, thereby predicting the video classification result. The present invention designs a temporal convolution CTL module with a temporal modeling mechanism and low computational complexity. CTL does not need to additionally model the time axis to integrate spatiotemporal features and motion features into a 2D convolutional network, which has the ability of temporal and spatial modeling while reducing the computing power cost. In addition, a model training method of pre-training plus fine-tuning is adopted. First, a video basic model is trained using an unsupervised method, and then the model is used as a pre-training model to perform fine-tuning training on the application scenario data. The training method can greatly improve the recognition accuracy of the model under the condition that the amount of training data is not large enough, so as to meet the application requirements of the anti-tailing scenario.

Claims

1. A video recognition-based anti-tailgating detection method, applied to mobile devices, characterized in that: The following steps are involved: Step A: Input a video image frame sequence and capture the video image using an equally spaced sampling strategy; Step B: During data preprocessing, the video image frame sequence is converted into the input format required by the video recognition network; Step C: Build a video recognition network model to achieve accurate recognition and real-time prediction of the model; The video recognition network is lightweight and uses time compression. The video recognition network includes a dimension conversion module, a convolution module, a maximum pooling module, four CTL modules, a downsampling module, an average pooling module, and a fully connected module. The CTL module consists of two branches: one branch is responsible for learning temporal importance, and the other branch with temporal position recovery capability can enhance the object modeling capability across time; The CTL module is a residual structure, including 1x1 convolution, CTL module and 1x1 convolution; The CTL module has a dual-branch structure, and the output of the CTL module is the sum of the output feature vectors of the two branches; The first branch of the CTL module is composed of TFC, i.e., temporal focused convolution. TFC introduces an attention mechanism to calculate the weight of each feature channel and obtains a feature vector with temporal importance. The second branch of the CTL module is composed of IOI cross-temporal interactions. IOI restores temporal position dynamics and temporal object relationships. The IOI module consists of two branches. One branch uses TFC to reduce the number of channels C to the number of frames T, while capturing the temporal importance, injecting temporal position coding information, and restoring temporal dynamics. It uses 7 × 7 convolution to model the object relationship between T frames, and uses 3 × 3 convolution to map the number of channels from T to C. The activation function is used to obtain the weight vector in the two-dimensional space of time and position. The other branch uses 3 × 3 convolution to output the feature map of C channels, and performs corresponding element multiplication with the weight vector to obtain a feature vector with temporal position dynamics and temporal object relationship. Step D: Define the classification loss function to supervise the model training. During the training process, the model gradually reaches a convergence state through the iteration of weight parameters. During the training, the optimal network weight parameters in the validation set are selected and saved to generate a video recognition model. Step E: In the actual reasoning process, directly load the weight parameters of the video recognition network model to complete the reasoning process of the model. In the reasoning mode, sample the input anti-tailing surveillance video to obtain a video image frame sequence, perform image preprocessing, and input the preprocessed data into the video recognition model for prediction. If the recognition result is normal, no alarm is triggered and the video stream continues to be input for monitoring. If the recognition result is tailing, an alarm signal is triggered.

2. The anti-tailgating detection method according to claim 1, characterized in that: The step A includes sub-step A1: capturing a frame of image at a fixed frame interval, where the frame interval ranges from 4 to 8, and the t-frame image sequence input to the video recognition network is recorded as .

3. The anti-tailgating detection method according to claim 2, characterized in that: The step B includes sub-step B1: converting the input format including resizing, standardization, and dimension conversion.

4. The anti-tailgating detection method according to claim 1, characterized in that: The step D includes sub-step D1: in the training mode, each training video is parsed into an image sequence and a labeling action is performed. The labels are divided into two categories: normal and trailing. The labeling rule is: when there is no one or only one person in the video, it is marked as normal; when the number of people in the video is greater than one person, it is marked as trailing.

Citation Information

Patent Citations

  • Airborne multi-sensor single-processor video compression processing method and device for unmanned aerial vehicle

    CN111447409A

  • Human body behavior recognition method and system based on extracted video spatio-temporal information

    CN113887419A

  • Video target segmentation method based on self-supervision

    CN114863348A

  • Video identification detection model generation method and device and computer equipment

    CN115424158A

  • Video anti-trailing method based on abnormal event detection

    CN118314530A